⚡ electrical engineering

Latency-Aware Deep Learning Benchmark for Real-Time Cyber-Physical Attack and Fault Classification in Inverter-Dominated Power Grids

This paper introduces a latency-aware benchmarking framework that evaluates eight deep learning architectures for real-time fault and cyber-attack classification in inverter-dominated power grids, revealing that while models achieve sub-cycle decision times, their end-to-end inference latency (50–90 ms) currently exceeds protection-grade requirements, highlighting a critical gap between algorithmic performance and deployable hardware solutions.

Original authors: Emad Abukhousa, Saman Zonouz, A. P. Sakis Meliopoulos

Published 2026-05-19

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Emad Abukhousa, Saman Zonouz, A. P. Sakis Meliopoulos

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the electrical grid as a massive, high-speed orchestra. In the past, the musicians (power generators) were predictable and steady. But today, the orchestra has added many new, digital instruments (solar panels and batteries) that play very differently. At the same time, a mischievous conductor is trying to sneak in and change the sheet music (cyber-attacks) or break a violin string (physical faults).

The goal of this research is to build a super-fast "conductor's ear" using Artificial Intelligence (Deep Learning) that can instantly hear a mistake and shout, "Stop! That's a broken string!" or "Wait! That's a fake note!" before the whole orchestra crashes.

Here is how the researchers tested this idea, broken down into simple steps:

1. The Training Ground (The Simulator)

You can't test a new conductor on a real, live orchestra without risking a disaster. So, the researchers built a hyper-realistic digital twin of a power grid using a professional simulator.

The Setup: They created a tiny, perfect version of a modern grid with solar panels and batteries.
The Data: They recorded the "music" (voltage and current waves) at a speed so fast it's like listening to every single vibration of a guitar string. They then injected 17 different types of "mistakes," ranging from a shorted wire to a hacker trying to trick the sensors.

2. The Contestants (The AI Models)

The researchers invited eight different types of "AI brains" to a competition to see who could identify these mistakes the best.

The Lineup: They ranged from simple, fast thinkers (like a basic calculator) to complex, deep thinkers (like a super-advanced detective).
The Test: They fed these AIs a continuous stream of data, just like a real-time radio broadcast, rather than a static recording.

3. The Results: The "Brain" vs. The "Body"

The competition revealed a fascinating split between how smart the AI is and how fast its body can move.

The Good News (The Brain is Fast): When it came to thinking about the data, the AI was incredibly sharp. It could look at the waveforms and figure out exactly what was wrong in less than one "cycle" of electricity (about 16 milliseconds). It was like a grandmaster chess player who sees the winning move instantly.
The Bad News (The Body is Slow): However, there was a catch. Even though the AI knew the answer quickly, the process of actually delivering that answer to the grid took much longer—about three to five times longer than the "one-cycle" safety limit required for real-world protection.
- Analogy: Imagine a race car driver who spots a turn and knows exactly how to steer (the AI's brain) in 10 milliseconds. But, because the car's steering wheel is heavy and the brakes are sticky, it actually takes 50 milliseconds to physically turn the car. In a real crash, that extra delay is too long.

4. The "Confidence" Safety Net

To make sure the AI didn't panic and shout "Fire!" when there was just a candle, the researchers added a "confidence filter."

How it works: If the AI isn't 100% sure of its answer, it stays silent (abstains) rather than guessing.
The Result: This prevented the system from making false alarms. If the AI saw a weird signal it had never seen before (like a fault in a different part of the grid), it would say, "I don't know what this is," instead of making up a wrong answer. This is crucial for safety; it's better to be unsure than to be confidently wrong.

5. The Verdict

The study concludes that while Deep Learning is smart enough to detect power grid attacks and faults with incredible accuracy, it is currently too slow to be the primary "emergency brake" for the grid.

Current Status: The technology is ready to be a "co-pilot" or a "situational awareness" tool that helps human operators understand what is happening.
The Gap: To become the main safety system that automatically shuts down the grid to prevent blackouts, the system needs to be optimized and run on faster, specialized hardware to close the gap between "thinking fast" and "acting fast."

In short: The AI has the right answers, but it needs a faster car to deliver them in time to save the day.

Technical Summary: Latency-Aware Deep Learning Benchmark for Real-Time Cyber-Physical Attack and Fault Classification

Problem Statement

Modern power systems are undergoing a transformation driven by the integration of Inverter-Based Resources (IBRs) and sophisticated cyber-physical infrastructure. This shift introduces complex failure modes where IBRs exhibit fault current characteristics (limited magnitudes, high-frequency harmonics) that challenge legacy protection schemes, while digitalization expands the cyber-attack surface. Although Deep Learning (DL) has shown promise in detecting physical faults and cyber-attacks with high offline accuracy, a critical gap remains in the literature: most studies prioritize accuracy on static datasets while neglecting end-to-end inference latency in realistic streaming contexts.

The operational timeline for power system protection is exceptionally demanding, often requiring fault isolation within one to two cycles (16–34 ms). Current research lacks rigorous, comparative benchmarks that quantify the latency of DL models against these stringent timing requirements. Consequently, claims of "real-time" applicability are frequently unsubstantiated, creating a barrier to deploying machine learning methods from research prototypes to protection-grade systems.

Methodology

To address this gap, the authors developed a latency-aware benchmarking framework using high-fidelity, time-domain signals generated from an industry-grade electromagnetic transient simulator (WinIGS).

1. Data Generation and Testbed:

Simulator: The dataset was generated using WinIGS to model grid dynamics, renewable integration, and anomaly injections with microsecond precision.
Testbed: An inverter-rich microgrid featuring a 50 kVA PV inverter and a 30 kVA Battery Energy Storage System (BESS).
Data Acquisition: Emulated IEC 61850 Sampled Value (SV) streams at 4.8 kHz (208 µs resolution) from two Merging Units (MUs).
Anomalies: The study covers 17 distinct classes, including normal operation, various physical faults (SLG, LL, DLG, Three-Phase), and cyber-attacks (CT/PT ratio attacks, GPS spoofing).
Preprocessing: Features were standardized, and sequences of 50 samples were generated using a sliding window approach.

2. Model Architectures:
Eight distinct DL architectures were systematically evaluated:

Multi-Layer Perceptron (MLP)
1D Convolutional Neural Network (1D-CNN)
Long Short-Term Memory (LSTM)
CNN-LSTM Hybrid
Transformer
Temporal Convolutional Network (TCN)
ResNet1D
Spatio-Temporal Graph Convolutional Network (ST-GCN)

3. Real-Time Inference Pipeline:
The evaluation was conducted on a streaming pipeline processing raw time-domain samples. To ensure decision stability, the framework employed:

One-Cycle Moving Average: A centered moving average filter ( $N_{cyc} = 80$ samples) to smooth probability vectors.
Confidence-Based Abstention: A decision mechanism where predictions are emitted only if the maximum confidence exceeds a threshold ( $\tau = 0.6$ ). Otherwise, the system abstains (outputs -1), preventing low-confidence misclassifications.

4. Evaluation Metrics:
The study defined four key metrics to assess performance under real-time constraints:

Accuracy: Proportion of correctly classified samples.
Coverage: Fraction of samples with confident (non-abstained) decisions.
Classification Time ( $T_{cls}$ ): The interval between anomaly onset and correct identification.
Inference Latency ( $T_{inf}$ ): The computational delay (average forward-pass time) required to generate a decision from raw measurements.

Key Contributions

The paper presents three primary contributions:

Systematic Comparison: A head-to-head evaluation of eight DL architectures on high-fidelity streaming data covering 17 different physical and cyber anomalies.
Practical Inference Pipeline: Implementation of a real-time pipeline featuring cyclic filtering and confidence-based abstention to enhance decision stability in streaming environments.
Quantitative Latency Benchmark: The first comparative evidence of sub-cycle detection and classification latencies for this diverse suite of models, explicitly distinguishing between algorithmic classification speed and end-to-end inference latency.

Results

Training Performance:
All models achieved high offline accuracy (>94%). The MLP model performed best with 99.66% accuracy despite having the fewest parameters (157k). The Transformer model demonstrated the highest balanced accuracy (99.88%), indicating robustness to class imbalance, though it had lower raw accuracy (95.37%).

Streaming Inference and Latency:

Stability: Raw streaming predictions exhibited significant jitter and oscillation. The application of the one-cycle moving average filter with confidence gating successfully stabilized classifications, eliminating false alarms and ensuring sustained model confidence.
Classification Time: All models achieved sub-cycle classification times (ranging from 3.38 ms to 14.83 ms), successfully identifying anomalies within the required protection window (typically <16 ms).
Inference Latency: Despite fast classification times, the end-to-end mean inference latency consistently exceeded three cycles (50–90 ms) across all models. For instance, the MLP had a latency of 57.20 ms, while the ST-GCN reached 88.89 ms.
Specific Scenarios:
- CT Ratio Attacks: MLP and Transformer models reacted fastest (<10 ms), while recurrent models (LSTM, CNN-LSTM) were slower (>50 ms).
- Out-of-Zone Faults: The system correctly identified faults outside the training zone by lowering confidence and flagging them as "unseen" (-1) rather than misclassifying them.

Significance and Claims

The paper concludes that while deep learning models possess the algorithmic capability to detect and classify power system anomalies with sub-cycle response times, a critical gap persists between this capability and protection-grade deployment. The consistent end-to-end inference latency (50–90 ms) exceeds the typical 16–34 ms window required for primary protection functions.

The authors assert that these findings establish a reproducible benchmark for sub-cycle anomaly detection. The work highlights that current latency profiles are more suitable for control and situational awareness functions rather than primary protection. The paper emphasizes that transitioning these methods to real-world applications requires further optimization and hardware acceleration to close the latency gap, rather than solely focusing on model accuracy. The study serves as a guide for understanding the trade-offs between accuracy, speed, and complexity in the design of next-generation intelligent protection systems.